Skip to content

MicroStructFormer

MicroStructFormer is a deep-learning–based framework for quantitative analysis of subcellular and periaxonal ultrastructures in Electron Microscopy images of myelinated fibers.
The project focuses on the systematic modeling of mitochondria-associated, periaxonal, and pathological compartments, enabling reproducible ultrastructural analyses across diverse experimental conditions.

Rather than focusing on conventional axon–myelin segmentation, MicroStructFormer is designed to provide a fine-grained semantic segmentation of mouse electron microscopy cross-sections, with particular emphasis on subcellular, periaxonal, and pathological structures that are difficult to capture with conventional pipelines.

Highlights

  • Targeted segmentation of mitochondria-related and periaxonal structures
    The framework prioritizes the detection and classification of ultrastructural components that are challenging to model with conventional pipelines, including mitochondria within axons, mitochondria outside axons, mitochondria-like organelles, periaxonal space, and morphologically abnormal or non-compact structures.

  • Transformer-based modeling of complex ultrastructure
    State-of-the-art transformer architectures (e.g. Mask2Former with Swin-Transformer backbones) are employed to capture heterogeneous morphologies and long-range contextual cues characteristic of subcellular and pathological features in EM data.

  • Quantitative morphometrics beyond pixel-wise segmentation
    MicroStructFormer supports downstream extraction of per-image and per-instance metrics, including area fractions, spatial distributions, compartment-specific statistics, and coupling relationships between organelles and myelin-associated geometry.

Scope and limitations

MicroStructFormer is optimized for transmission Electron Microscopy datasets of mouse fibers and is intended as a research framework rather than a general-purpose EM segmentation tool or a biophysically explicit simulator.
In the current implementation, axon and myelin areas are provided by external tools such as AxonDeepSeg, and the accuracy of downstream analyses depends on the quality of these external segmentations. Performance may further degrade in cases involving severe membrane disruption, ambiguous compartment areas, or domain shifts beyond the training distribution.

  • Architecture
    Model components, target structures, and architectural design choices.

  • Dataset
    Our dataset references and annotations on Roboflow.

  • Model Zoo
    Pre-trained model weights, example configurations, and post-processing scripts.

  • Citation
    How to cite MicroStructFormer and its external dependencies in academic publications.